Safety correlation and implications of an in-vehicle data recorder on driver behavior
Oren Musicant
Department of Industrial Engineering and Management
Ben Gurion University of the Negev, Israel
Tel: (972) 3 6340238 Fax: (972) 3 6340397 Email: [email protected]
Tsippy Lotan
OR YAROK,
38 Hashoftim St., Ramat Hasharon 47210, Israel,
Tel: (972) 3 760-6162 Fax: (972) 3 760-6160 Email: [email protected]
Tomer Toledo
Department of Civil and Environmental Engineering,
Technion – Israel Institute of Technology, Haifa 32000, Israel
Tel: (972) 4 829-3080 Fax: (972) 4 829-5708 Email: [email protected]
Word count: Tables and figures
12 x 250 =
3000
Word count
4443
Total
7443
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
ABSTRACT
Understanding individual on-road behavior has been a key issue in traffic safety research. A
deeper insight into the driving patterns has been so far restricted, due to limited availability of
actual longitudinal data and exposure data. This paper presents an innovative approach in
which actual driving data is obtained by means of an in-vehicle data recorder (IVDR). This
data are used to evaluate drivers' behavior and safety. In particular, two major issues
regarding the validity of such systems as safety promoters are analyzed. The first is the
correlation between the accident risk indices calculated based on the IVDR data and drivers'
records of car crash involvement. Once this correlation is established, the second issue is the
ability of drivers to change their behavior in the desired direction, namely to become safer
drivers. The results obtained, based on observations of 103 drivers in 18173 trips and more
than 8400 driving hours indicate that the risk indices and classifications calculated based on
the IVDR data are significantly correlated with past car crash involvement. Exposure to the
feedback generated from the system has a potentially high impact on collision reduction with
over 40% reduction in crash rates using before and after data. Finally, this behavioral change
has been maintained for 9 months after the exposure to the IVDR feedback.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
INTRODUCTION
Human behavior and driver errors are a cause in the overwhelming majority of road crashes
[1]. Aside from a few general parameters such as age, gender, time of day and type of
collision, good understanding of collision cause and individual drivers' risk levels are limited.
Even when crashes are thoroughly analyzed – the analysis refers mostly to the circumstances
of the collision itself and not to patterns of the behavior of the drivers involved. Most
methods used to define drivers' skills and styles are based on stated preference tools, which
could be biased and general. Otherwise, evaluations of driving behavior may only be
performed on a limited scope and scale, for example in experiments involving a driving
simulator [2]. Recently, with the advances in monitoring technologies, a new generation of
data collection approaches is becoming available to evaluate actual driving behavior
continuously, in much more detail and with large scale implementations. In-vehicle data
recorders (IVDR), which are electronic monitoring devices installed in the vehicle, are one of
several measures that enable to accurately measure drivers' behavioral patterns and to
evaluate various strategies to influence drivers' behavior. This paper describes an evaluation
of the potential of one such IVDR system to serve as a tool to evaluate and manage drivers'
risk of crash involvement.
The rest of this paper is organized as follows: the next section reviews literature on
IVDR systems and their potential uses and impacts. We then describe the IVDR system used
in this study and report on results of experiments designed to validate its measurement and to
evaluate its potential to impact drivers' behavior. The final section summarizes our findings
and discusses potential future applications.
IN-VEHICLES DATA RECORDERS
In-vehicle data recorders (IVDR) are on-board devices that record information about the
movement, control and performance of the vehicle [3]. IVDR technology is generally based
on the collection of data on the vehicle performance and its immediate environment. This
data is analyzed to detect patterns, which are then matched with decision rules that control
activation of certain driver support functions [4]. A number of IVDR systems have been
developed in recent years. While their details, capabilities and intended use vary, the
information they commonly collect may be classified in several categories ([5], [6]):
1. Vehicle movement: i.e. longitudinal and lateral accelerations and the speed of the
vehicle.
2. Driver control, which includes variables such as the engine throttle and brake
application and wheel-angle.
3. Engine parameters, such as RPM.
4. State of the vehicle safety systems, such as air bags, seat belts, ABS and traction
control.
5. Vehicle location, using GPS systems.
6. Time.
7. Visual documentation both inside and outside the vehicle.
Safety-related applications of IVDR systems can be broadly classified into two categories:
1. Short-term prevention, such as with speed adaptation, collision avoidance systems,
weather information, vision enhancement and lane keeping assistance.
2. Analysis of collision events, such as within collisions investigations, emergency
response and research and development of safety devices.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
In addition to these applications, feedback and training based on the information
collected by IVDR may also help reduce collision risk in the longer-term [7]. For example,
Lehmann [8] reports several case studies in which the installation of IVDR systems in
various fleets resulted in reductions of 20-30% in collisions rates, and even more significant
reductions in the related costs. Similar reduction rates were reported in an experiment
reported by Wouters and Bos [9]. Several on-going studies aim to develop and evaluate
IVDR systems that will provide feedback to drivers. Some systems were intended for the
insurance market (e.g. [10]), while others serve research needs ([11], [12]). NHTSA [13] has
recently conducted an ambitious study in which 100 vehicles were instrumented with IVDR
as well as video cameras, radar sensors, GPS and lane trackers for a period of 13 months. The
data collected with these vehicles include a full documentation of 82 crashes, 762 near
crashes and more than 42000 hours driven. Preliminary analysis of the huge data set
collected in this study indicates great potential to enrich traffic safety research such as
evaluation of risky driving behavior and collision risk, evaluation of relative risk of engaging
in secondary tasks, and evaluation of driver response to lead vehicle brake lights. These
results illustrate the capabilities of the technology to evaluate behavior and increase safe
driving. In summary, IVDR hold a potential for significant improvements in traffic safety,
however, further research is needed to validate these systems and to identify the data
collection needs and feedback types that would best serve this goal.
THE DRIVEDIAGNOSTICS IVDR SYSTEM
The specific IVDR system evaluated in this paper was developed by DriveDiagnostics LTD.
The overall system framework is described in Figure 1. The system performs a sequence of
four data collection and analysis steps: measurement, identification, analysis, and reporting.
1. Measurement. In this module information on the movement of the vehicle is
collected. This information includes the two-dimensional acceleration and speed of
the vehicle at a sampling rate of 40 measurements per second. The system also
records the position of the vehicle using GPS.
2. Detection. In this module pattern recognition algorithms are applied to the raw data in
order to detect certain maneuvers that the vehicle performs. The system currently
identifies over 20 different maneuver types, such as lane changes, sudden brakes,
strong accelerations, excessive speed and so on. The identified maneuvers are also
classified in terms of their severity based on parameters of the detailed trajectory of
the vehicle during the maneuver, such as the maneuver duration, extent of sudden
changes during the maneuver, the speed it is performed at and the magnitude of the
forced applied on the vehicle.
3. Analysis. The information on the maneuvers detected by the system is stored in a
database. The analysis module uses this information to build several driver-specific
and vehicle-specific indices and statistics that can be used to characterize drivers'
behavior and their personal risk of crash involvement. The following indices are used
in the current implementation:
• Overall risk index: this index is a numeric measure that aims to estimate the
drivers' risk of involvement in car crashes. Values typically range between 0 and
500 with high values implying higher risk. The index depends on the types,
numbers and severity of maneuvers that the driver has performed. The
accumulated of driving time is used to normalize the amount of maneuvers.
• Risk classification: based on the risk index, drivers are classified in three risk
categories. One of the main purposes of this classification is to provide a simpler
system to report risk indices that will be understandable to layman drivers. For
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
•
•
•
•
this reason the classes are also labeled as "Green" (moderate behavior), "Yellow"
(Intermediate behavior) and "Red" (aggressive behavior).
Trip-level risk index and classification: a risk index is also calculated for each
individual trip. This index takes into account the maneuvers that occurred during
the trip and their severity as well as other factors such as the trip duration. As with
the overall index trips are also classified in three risk categories: Green, Yellow
and Red.
Speed index: this numeric value reflects the driver's speed profile. While it is also
taken into account in the overall risk indices, speed has been shown to be an
important predictor of crash risk involvement in many studies.
Driving Patterns indices: In addition to the speed index, other indices that reflect
drivers' performance in several specific categories are calculated using only the
information on the relevant maneuvers. Examples include performance indices for
corner handling, lane keeping and braking patterns.
Fuel consumption index: the information related to drivers' performance is
combined with other information such as vehicle types to predict fuel
consumption.
4. Reporting. The final module is concerned with providing feedback to drivers.
Feedback may be provided in a variety of forms and through several media channels.
Reports can be tailored to include various information at different levels of detail that
may be of interest to fleet safety managers, insurance companies, vehicle owners,
drivers and researchers. An example of a monthly driver report is shown in Figure 2.
In the figure, each square corresponds to a trip. The X axis indicates the day of the
month and the Y axis indicates the number of trips performed during each day. Trips
are color-coded by their classification: green, yellow and red for trips in which the
driver behavior was classified as moderate, intermediate and aggressive, respectively.
Detailed information on each trip and on the driving pattern indices in the various
subcategories is also provided. In addition, the report includes information on the total
hours of driving during the month and comparison of the drivers' performance to
previous months and to other drivers in the fleet. These reports may be delivered
through an internet website, emails or as printed reports. In addition, real-time
feedback on current trips may be provided through mobile phone text messages and
warning lights on an in-vehicle display board.
In terms of the communications systems, the first two modules are performed in real-time
within the unit installed in the vehicle. Only the detection outcomes are transmitted to the
application server. This design significantly reduces the amount of communications required
between the vehicle and the server. The information is transmitted in real-time, continuously
throughout the trip, through wireless networks. The analysis and reporting modules are
performed at the application server. The server also stores other relevant information, such as
collision records, maintenance and fuel costs etc.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
Measurement
Acceleration
Speed
Position
Detection
Maneuvers
Severity
Analysis
Risk index
Risk classification
Speed index
Driving Patterns
Costs
Reporting
Real time
Off line
Figure 1 Overall framework of the DriveDiagnostics system
Figure 2 An example of a monthly driver report
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
OBJECTIVES AND METHODOLOGY
The main hypothesis underlying the application of the IVDR system is that the data collected
and the risk indices calculated using these data are highly correlated with the actual risk of
car crash involvement and therefore can be used as indicators to that (latent) risk. Preliminary
results demonstrating the linkage between drivers' risk indices and their crash records were
presented by Lotan and Toledo [14]. However their analysis was based on evaluation of only
29 drivers. In this study we conduct a new evaluation of the connection between risk indices
and crash records using a larger and more diverse sample. After this connection is established
we also evaluate the potential usefulness of the system, through the feedback it provides to
drivers and other entities, to reduce crash rates and to sustain any reductions over time. We
next describe the experiment procedure and results.
Sample selection
The drivers in the sample were chosen from the fleets of 6 different organizations in Israel
that provide vehicles to their employees as part of their benefit programs. The drivers in the
sample are between 24 and 65 years old, with an average of 41 years. These drivers have
been using company vehicles for periods that ranged between 6 and 80 months, with an
average of 39 months. These drivers, who are not professional drivers, may use these vehicles
to commute, make other business trips and for any other personal trips. While the possibility
to identify drivers with personal codes is available, it was not implemented in these
commercial installations. Therefore, out of all the IVDR installed vehicles in these fleets, all
vehicles that were only driven by a single driver were selected for the sample. Descriptive
statistics of the resulting sample are summarized in Table 1.
Sample statistics
Drivers
Males
Females
103
80
23
Driving
hours
8430
Trips
18173
Maneuvers
detected
35168
Table 1 Sample descriptive statistics
Procedures and measurements
The implementation of IVDR systems in all vehicles in the six organizations followed a
similar process outlined in the following steps:
1. Initially, the systems are installed in the vehicles without providing a detailed
explanation to drivers about the nature of the system or any feedback from it. Privacy
protection laws dictate that the drivers consent to the installation, and so they were
notified, but only received a general explanation that these are safety-related
instruments that will be used in an experiment. The data collected in this initial
period, which we refer to as the "blind profile" period, is useful to characterize the
habitual behavior of these drivers prior to receipt of any feedback from the system.
2. About 8 weeks later, the participating drivers in each organization were invited to a
meeting in which they learned about the IVDR system, its working and the feedback
it provides. In this meeting they also received an initial feedback on their own driving
and an access code to a secured web site where they could access their own data, as
well as compare their performance to the fleet's averages. In all cases, it was
determined that the IVDR records will not be used in any way against the drivers.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
For each driver, all the data detected and analyzed by the IVDR system on all trips performed
in the vehicle was collected. From these the indices and classifications that were described
above were calculated. These are used as the independent variables on the analysis that
follows. In addition, four other variables that represent drivers' past involvement in car
crashes were collected from the records of the organizations that provide and maintain the
vehicles. These variables, which are used as the dependent variables in our analysis, are
defined as follows:
1. Collision rate, which is the number of crashes per year that the driver has been
involved in. Crashes were counted from the date that the driver received an
organization vehicle, but no more than 3 years before the date of IVDR installation.
2. Fault collisions rate, which is the number of crashes for which it was determined for
insurance purposes that the driver was at fault.
3. Collision cost rate, as estimated for insurance purposes. The cost is the sum per year
of the costs of all crashes that were recorded in New Israeli Sheqels (4.5 NIS ≈ 1$).
This costs id the direct cost of the vehicle repair. No other costs were included.
4. Annual fault collision cost, which is the cost per year of fault collisions.
RESULTS
In this section we present results of analysis based on the data collected in the experiment
described above.
Linkage of IVDR risk index and classification to crash history
As noted above, the risk index and classification computed by the IVDR systems may be
useful, if it can be shown that they are highly correlated with the actual risk of car crash
involvement. However, the risk cannot be directly measured and so we use past crash records
as indicators to the risk. Crash records were collected for the 103 drivers in the sample. The
IVDR risk indices used are the ones computed for the driving that these drivers undertook
during the blind-profile period, which refers to the time between the device installation date
and the first time the driver was exposed to the system data. Descriptive statistics of these
data are presented in Table 2. The zero median values reflect the fact that over 50% of the
drivers in the sample were not involved in car crashes in the 3 years prior to the installation of
the IVDR.
Mean
Median
Std. Deviation
Minimum
Maximum
Risk
index
40
28
38
2
206
Crash
rate
0.38
0.00
0.66
0.00
4.0
Fault
crash rate
0.20
0.00
0.41
0.00
2.00
Crash
costs
986
0
5188
0
46958
Fault
crash costs
826
0
5014
0
46958
Table 2 Risk indices and crash data descriptive statistics
Table 3 shows the average and median values of the four safety indicators described
above for drivers that were classified in each of the three risk categories: Green (moderate),
Yellow (Intermediate) and Red (aggressive). In the sample, 39 drivers were classified as
Green, 41 Yellow and 23 Red. With all crash indicators, the values are higher for the riskier
classes: Red compared to Yellow and Yellow compared to Green. The differences are more
significant for crash costs compared to numbers of crashes, which indicates that the crashes
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
that drivers in the riskier classes are involved in are also more severe. It should be noted that
there were no significant differences (p=0.582) between the driving hours of drivers in the
three categories.
Crash rate
Fault crash rate
Crash costs
Fault crash costs
Green
Average Median
0.18
0.00
0.09
0.00
270
0
194
0
Yellow
Average Median
0.34
0.00
0.14
0.00
515
0
371
0
Red
Average
0.79
0.49
3783
3219
Median
0.33
0.16
351
0
Table 3 Collisions and Fault collisions by driver safety class
In order to test the statistical significance of the difference among the risk classes, a
Kruskal-Wallis (K-W) test was conducted. This non-parametric test is designed to test the
null hypothesis that the median crash rates in the three categories are equal. It does not
assume normal distribution of the variables and therefore suitable for crash rate data [15].
The table reports the p-values for this analysis. The equality of medians can be rejected for
crash rates and fault crash rates at 95% confidence and for crash costs at 90% confidence.
The post hoc analysis is designed to identify homogeneous groups in the three categories. It
was conducted through multiple applications of Wilcoxon Rank-Sum test. p-values were
adjusted to the multiple categories using the Benjamin and Hochberg FDR rule. The analysis
grouped together the Green and Yellow categories to one homogenous group that differs
from the Red category for crash rates, fault crash rates and crash costs variables (p ≤ 0.05).
No significant differences were found among the categories with regard to the fault crash
costs. These results imply that the driver classifications are indicative of crash risk and costs
involvement. Table 4 summarizes the results of these tests.
Although the analysis is adequate to the non-normal distribution of the crash rates and
costs, it could be argued that the median does not characterize the safety categories as well as
the mean because it does not capture the long tail affect. We therefore also conduct analysis
on the mean values in the various categories. The sample averages and standard errors are
present at Figure 3. To test the null hypothesis of equality of means one-way ANOVA and
the Scheffe post hoc test were used. The results are also shown in Table 4. The equality of
means can be rejected for crash rates and fault crash rates at 95% confidence and for crash
costs and fault crash costs at 90% confidence. With all the indicators but the crash costs, the
Scheffe test shows that the Green and Yellow class can be grouped together as being
significantly different from Red drivers.
Statistic
K-W
p-value
Crash rate
0.001
Fault crash rate
0.007
Crash costs
0.068
Fault crash costs
0.236
Post Hoc
results
(green, yellow)
(red)
(green, yellow)
(red)
(green, yellow)
(red)
ANOVA
p-value
0.068
-
-
0.091
(green, yellow)
(red)
0.001
0.000
Scheffe Test
results
(green, yellow)
(red)
(green, yellow)
(red)
Table 4 Test results on the collisions and fault collisions by driver safety categories
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
Crashes per year
Fault Crashes per year
1.50
1.00
Error bars: +/-...
Error bars: +/-...
1.25
0.80
Average
Average
1.00
0.75
0.60
0.40
0.50
0.20
0.25
0.00
0.00
Green
Yellow
Red
Green
Safety Class
Yellow
Red
Safety Class
Crashes at fault costs per year
Crashes costs per year
7000
6000
Error bars: +/-...
Error bars: +/-...
6000
5000
5000
Average
Average
4000
4000
3000
3000
2000
2000
1000
1000
0
0
Green
Yellow
Safety Class
Red
Green
Yellow
Red
Safety Class
Figure 3 Average and standard deviation of the number of crashes and crashes costs per
year by driver class
The risk index, which is continuous, is intended to provide a higher level of resolution
on drivers' risk of crash involvement. To evaluate its compatibility with crash history linear
regression analyses were conducted. Although the Poisson regression is more common for
the analysis of crash rate data, linear regression was found more adequate since many of the
drivers in the sample were not involved in any car crashes. The model also takes into account
the existence of differences between the organizations in the approaches they use to record
and investigate crashes. Organization-specific dummy variables were introduced in order to
capture the organization effects. The resulting model is given by:
Yi = α j Dij + β X i + ε i
(1)
Yi and X i are the crash history variable and risk index for driver i, respectively. Dij
is a dummy variable, which takes the value 1 if driver i is with organization j, and 0
otherwise. α j and β are parameters. ε i is a random error term.
The variance of the risk index values calculated based on the IVDR measurements
decreases when more data is accumulated. Therefore, to account for this effect and maximize
the efficiency of the parameter estimators, weighted least squares regression was used.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
Weights were set according to the amount of driving each driver undertook during the blind
profile period. The regression results are presented in Table 5. The slope β of the regression
line captures the marginal effect of a change in the IVDR risk index on the predicted crash
rates. This impact is, as expected, positive in all cases. It is significantly larger than zero at
95% confidence for all four cases. Figure 4 demonstrated the impact of the IVDR risk index
on crash involvement as predicted by the regression model. The model predicts that a driver
with a 150 points risk index is 42% more likely to be involved in a car crash and 46% more
likely to be involved in own fault car crash compared to a driver with a 100 points risk index.
The associated costs per year are likely to be 59% and 56% higher, respectively.
R2
0.442
0.441
0.267
0.225
Dependent parameters
crash rate
fault crash rate
crash costs
fault crash costs
β (p-value)
0.007 (.000)
0.005 (.000)
105.2 (.004)
75.63 (.020)
1.8
1.6
1.4
1.2
20000
18000
16000
All
Fault
Crash Rate
Crash Rate
Table 5 Results of regression of crash rates on IVDR risk indices
1.0
0.8
0.6
0.4
0.2
0.0
All
Fault
14000
12000
10000
8000
6000
4000
2000
0
0
50
100
Risk Index
150
200
0
50
100
Risk Index
150
200
Figure 4 Impact of IVDR risk index on crash rates and crash costs
Impact of IVDR on collision reduction
After the blind profile period the drivers were exposed to the information collected by the
system and were able to access the reports generated for them as shown in Figure 2 through a
secured website. In all organizations, no other action was taken in order to affect drivers'
behavior. Thus, it can be assumed that changes in crash rates can be attributed to the feedback
received from the IVDR. To evaluate the impact of the IVDR a before and after study was
conducted, in which the crash rates in the two periods were compared. The period before
installation, refer to all the data that was available for the drivers in the past (from the time
they received the organization vehicle). The period after installation, refer to all the data that
was available for the drivers after they were exposed to the IVDR feedback. Data on the 70
drivers which used the IVDR system feedback for at least 5 months (and up to 14 months)
was used in this analysis. To make the data comparable all data were normalized to annual
rates and costs. In addition we found that there were no significant differences in the driving
hours among the blind profile and each of the months after the exposure to the IVDR
feedback, which indicate that drivers were not less exposed to risk. Figure 5 shows the
average and standard errors of crash rates and costs in the periods before and after the
exposure to the IVDR feedback. One-tailed t-tests were conducted in order to test the
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
statistical significance of the reductions in crash rates and costs that were observed. The test
results are presented in Table 6. The results indicate reductions of 44% (p=0.02) in collisions
rates and 38% (p=0.097) in fault collisions rate. The decreases in costs are even higher and
statistically significant at the 95% confidence level. The reduction in crash rates reported here
is more substantial compared to those between 20% and 30% reductions reported by
Lehmann [8]. A possible explanation to some of the difference may be that Lehman's study
included drivers of police vehicles and buses, which may not be as sensitive to the feedback
as other drivers.
t-statistic
2.086
1.309
2.411
1.864
Average crash rate
Average fault crash rate
Average crash costs
Average fault crash costs
p-value
0.020
0.097
0.010
0.034
Table 6 Collisions and Fault collisions before and after IVDR installation
Crashes
0.6
Crashes at fault
0.30
Error bars: +/- 1.00 SE
Error bars: +/- 1.00 SE
0.25
0.4
Average
Average
0.5
0.3
0.2
0.20
0.15
0.1
0.10
0.0
Before
After
Before
Crashes Costs
Crashes at fault costs
700
Error bars: +/- 1.00 SE
800
600
600
500
Average
Average
After
400
Error bars: +/- 1.00 SE
400
300
200
200
100
0
0
Before
After
Before
After
Figure 5 collisions rate and costs before and after exposure to IVDR feedback
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
Sustainability of the IVDR impact
While the results indicating that the feedback from the IVDR system can have a substantial
impact on safety are important, it may only be a short-term impact. At present there is not
enough longitudinal crash data to fully evaluate this issue. However, an evaluation of the
temporal effects could be conducted through a comparison over time of the risk indices,
which have been correlated with the risk of crash involvement. For that purpose the risk
indices were calculated for the blind profiling period and for each one of the 10 months
following the exposure to the system. Figure 6 shows the average and 85th percentile of the
risk indices distribution as well as the average speed index for the various months. The
reduction in the average risk index from the blind profile to the first month after the exposure
is 27%. Furthermore, there is a reduction of 30% in the 85th percentile risk index, which
implies that the moderating effect is obtained not only on average but also for the drivers with
highest risk indices in the sample. A similar reduction (32%) is also observed when
considering only the speed component of the risk indices. The differences between the blind
profile period and each one of the 9 months that follow in the average risk index are
statistically significant at the 90% confidence level. The 10th month average risk index is not
significantly lower than that of the blind profile period.
100
90
80
Risk index
70
60
50
40
30
20
10
0
Blind
Profile
1
2
3
4
5
6
7
8
9
10
Months since exposure
85th Percentile
Average
Speed index
Figure 6 Temporal changes in risk indices statistics
DISCUSSION
This paper describes the overall framework and the components of an IVDR system called
DriveDiagnostics, and presents results from a study to evaluate its diagnostic validity and
impact on a driver’s behavior. The system records the movement of the vehicle and uses this
information to detect and classify over 20 different maneuver types. These maneuvers are
then used to build a driver profile by calculating several risk indices and classifications.
Three of those indices; overall driving risk index and classification and the speed index were
used in this paper to investigate several aspects of the IVDR performance. Firstly, a
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
significant link between the IVDR indices to drivers' past crash records history was
established. This result suggests that the driving risk indices calculated by the system can be
used as indicators to the risk of involvement in car crashes at the level of the individual
driver. A study of crash rates in the period before and after drivers' were exposed to the
feedback from the IVDR system shows statistically significant reductions of over 40% in the
crash rates. The cost reduction was also prominent, but not as statistically significant. It
should be noted that these results are based on a relatively short period of time after the
exposure to the system and therefore need to be further evaluated. Finally, an evaluation of
the temporal changes in the IVDR risk indices showed that the decrease in the average and
85th percentile of the risk indices distribution and in the average speed index are sustained
over the first 9 months after the exposure to the IVDR feedback. In summary, the results
presented in this paper demonstrate that the IVDR data are significantly correlated with the
risk of crash involvement and that the feedback provided based on these data is useful in
reducing this risk. The results thus suggest that an IVDR system that can monitor drivers’
behavior and produce statistics that indicates on their safety performance may be a useful in
several applications and research studies related to driving behavior and safety. Further
research is still required to provide deeper the understanding of the IVDRs long term affect
and the its potential application with specific groups of drivers (e.g. young drivers,
professional drivers).
REFERENCES
1. Evans L. (2004) Traffic safety, Science Serving Society, Bloomfield Hills, MI.
2. Comte, L.S, (2000) new systems: new behaviour? Transportation Research Part F, 3,
pp 95-111
3. Correia J.T., Iliadis K.A., McCarron E.S. and Smolej M.A. (2001) Utilizing data from
automotive event data recorders, Proceedings of the Canadian Multidisciplinary Road
Safety Conference XII, London Ontario.
4. Van der Heijden R. and Van Wees K. (2001) Introducing advanced driver assistance
systems: some legal issues, European Journal of Transport and Infrastructure
Research, 1, pp. 309-326.
5. Chidester A.C.D, Hinch J. and Roston T.A. (2001) Real world experiences with event
data recorders, Proceedings of the 17th International Technical Conference on the
Enhanced Safety of Vehicles (ESV), paper 247.
6. NHTSA (2001) Event data recorders: summary of findings by the NHTSA EDR
working group, Report NHTSA-99-5218, Department of Transportation, Washington
DC.
7. U.S. Department of Transportation (2001) A report to congress on electronic control
module technology for use in recording vehicle parameters during a collision, Report
DOT-MC-01-110, Washington DC.
8. Lehmann G. (1996). The features of the accident data recorder and its contribution to
road safety. Proceedings of the 15th International Technical Conference on Enhanced
Safety of Vehicles (ESV) Volume 2, pp. 1565-1568.
9. Wouters P.I.J. and Bos J.M.J. (1997) The impact of driving monitoring with vehicle
data recorders on accident occurrence: methodology and results of a field trial in
Belgium and the Netherlands, Report R-97-8, SWOV, Dutch Institute for Road Safety
Research.
10. TripSense. Web-Page: https://tripsense.progressive.com/home.aspx, Accessed August
1, 2005.
11. Boyce T. (2001) A technology to measure multiple driving behaviours without self
report or participant reactivity, Journal of Applied Behavior Analysis, 34, pp 39-55.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.
12. DriveAtlanta (2002) DriveAtlanta: study will monitor driver behavior to determine
role of speed in highway crashes, Georgia Tech Research News.
http://www.gtresearchnews.gatech.edu/newsrelease/DRIVEATL.htm. Accessed
August 1, 2005.
13. Neale V.L., Klauer S.G., Knipling R.R., Dingus T.A., Holbrook G.T. and Petersen A.
(2002) The 100 car naturalistic driving study phase I: experimental design, Report
DOT-HS-808- 536, Department of Transportation, Washington DC.
14. Lotan T. and Toledo T. (2006) An in-vehicle data recorder for evaluation of driving
behavior and safety. Preprints of the 85th Annual Transportation Research Board
Annual Meeting.
15. Tesema T.B., Abraham A. and Grosan C. (2005) Rule mining and classification of
road traffic accidents using adaptive regression trees. International Journal of
Simulation systems, 6, pp 80-94.
TRB 2007 Annual Meeting CD-ROM
Paper revised from original submittal.